The privacy-first knowledge fabric that turns every law firm's accumulated expertise into a queryable, governable, pluggable intelligence layer.
Trellis is a two-layer knowledge platform for law firms. It captures individual lawyer expertise locally, governs its publication into a shared team knowledge graph with AI-assisted redaction, and makes the accumulated intelligence queryable via a citation-grounded chatbot with a signature graph-overlay visualization.
The problem: Law firms lose an estimated $15M–$40M annually to knowledge fragmentation, redundant work, and institutional memory loss when senior lawyers depart. Attorney-client privilege structurally blocks adoption of generic AI tools.
The solution: A personal brain (local, private) where lawyers capture thinking throughout the day, paired with a team-managed knowledge graph (governed cloud) where sanitized insights are published and retrievable by the entire practice group.
| Feature | Description |
|---|---|
| Multimodal Capture | Text, audio (Whisper transcription), and image (Gemini Vision OCR) note intake |
| AI Auto-Organization | Gemini extracts entities, classifies notes, and builds a personal knowledge graph |
| Privacy-First Publishing | Two-pass redaction pipeline (Presidio PII tokenization + Gemini generalization) with side-by-side diff review |
| Team Knowledge Graph | Governed, queryable graph of the practice group's accumulated intelligence |
| Citation-Grounded Chat | RAG-powered retrieval over the team graph; every claim cites a source node |
| Query-Overlay Visualization | Signature visual moment — chat dims, team graph fades in, cited nodes pulse as the answer streams |
| Pluggable Brain (V1) | MCP server endpoint so external AI tools (Harvey, CoCounsel, Copilot) can query the firm's knowledge |
| Layer | Technology |
|---|---|
| Frontend | React 18, Vite, TypeScript, Tailwind CSS, Cytoscape.js |
| Backend | Node.js 20, Express, TypeScript |
| Database | PostgreSQL 16, pgvector |
| AI | Gemini 2.5 Pro/Flash/Vision, Whisper API, Microsoft Presidio |
| Personal Storage | IndexedDB (browser) |
| Deployment | Vercel (frontend), Railway/Render (backend + Postgres) |
trellis/
├── apps/
│ ├── web/ # React + Vite frontend (SPA)
│ │ └── src/
│ │ ├── api/ # Backend client (fetch wrappers, TanStack Query)
│ │ ├── components/ # Reusable UI components
│ │ ├── lib/ # IndexedDB wrappers, graph utilities
│ │ ├── store/ # Zustand stores
│ │ ├── styles/ # Design tokens, global CSS
│ │ └── views/ # Page-level views (auth, capture, chat, graph, publish, team)
│ └── api/ # Node.js + Express backend
│ └── src/
│ ├── routes/ # Express route handlers
│ ├── services/ # AI orchestration, redaction, RAG
│ ├── db/ # Postgres client, queries, migrations
│ ├── prompts/ # System prompts for Gemini
│ └── seed/ # Seed data scripts and content
├── infra/
│ ├── docker-compose.yml # Local dev: Postgres + Presidio
│ └── deploy/ # Vercel + Railway deployment configs
├── docs/ # Hackathon submission assets, public docs
├── .agent/ # Agent-facing specification documents
│ ├── product-brief.md # Product vision, market, business model
│ ├── product-requirements.md # PRD: roles, features, acceptance criteria
│ ├── project-architecture.md # System architecture, data model, pipelines
│ ├── design-guidelines.md # Brand, color, type, motion, components
│ ├── context-dump.md # Full decision history and reasoning
│ └── trellis-vault-assistant.md # Vault assistant system prompt
├── vault/ # LLM-maintained knowledge wiki (Obsidian vault)
│ ├── sources/ # Summary pages per source document
│ ├── entities/ # Named things: products, tools, companies, roles
│ ├── concepts/ # Ideas, patterns, doctrines, techniques
│ ├── topics/ # Synthesis pages spanning multiple sources
│ ├── raw/ # Immutable source documents
│ └── templates/ # Page templates for wiki maintenance
├── AGENT.md # Agent entry point (start here)
└── README.md # This file
Initial wedge: Litigation practice groups at mid-size law firms (50–300 lawyers).
Expansion path: Other practice groups → BigLaw / AmLaw 200 → In-house legal teams → Adjacent professional services.
Business model: Practice-group license, tiered by size ($25K–$85K ARR), with land-and-expand across the firm (~$200K ARR at maturity).
Built for the AI & Big Data Expo Lablab Hackathon — Track 4: Data & Intelligence. Competing for the Gemini Award (Gemini powers extraction, redaction, and synthesis throughout).
Demo narrative (5 min): Capture → Publish with visible redaction → Retrieval with graph-overlay visualization.
 Keith Ruezyl |
 Gabe San Diego |
 Nicolo Porter |
This project was built for the AI & Big Data Expo Lablab Hackathon.